1. Challenging the Myth of Graph Collaborative Filtering:
a Reasoned and Reproducibility-driven Analysis
Vito Walter Anelli1, Daniele Malitesta1, Claudio Pomo1,
Alejandro Bellogín2
, Eugenio Di Sciascio1
, Tommaso Di Noia1
Politecnico di Bari, Bari (Italy), email: firstname.lastname@poliba.it
Universidad Autónoma de Madrid, Madrid (Spain), email: alejandro.bellogin@uam.es
The 17th ACM Conference on Recommender Systems
Singapore, SG, 09-20-2023
Main Track - Reproducibility
1
2

2. Challenging the Myth of Graph Collaborative Filtering: a Reasoned and Reproducibility-driven Analysis
The 17th ACM Conference on Recommender Systems (Singapore, 18-22 September 2023)
● Introduction and motivations
● Background and reproducibility analysis
● Replication of prior results (RQ1)
● Benchmarking graph CF approaches using alternative baselines (RQ2)
● Extending the experimental comparison to new datasets (RQ3 - RQ4)
● Conclusion and future work
Outline
2

3. Introduction and motivations
3

4. Challenging the Myth of Graph Collaborative Filtering: a Reasoned and Reproducibility-driven Analysis
The 17th ACM Conference on Recommender Systems (Singapore, 18-22 September 2023)
In collaborative filtering (CF), graph convolutional networks (GCNs) have gained momentum thanks to their ability to
aggregate neighbor nodes information into ego nodes at multiple hops (i.e., message-passing), thus effectively
distilling the collaborative signal
One-hop
neighborhood
m
e
s
s
a
g
e
message
m
e
s
s
a
g
e
m
e
s
s
a
g
e
4
Graph collaborative filtering: message-passing

5. Challenging the Myth of Graph Collaborative Filtering: a Reasoned and Reproducibility-driven Analysis
The 17th ACM Conference on Recommender Systems (Singapore, 18-22 September 2023)
Graph collaborative filtering: a non-exhaustive timeline
2020
Lighten the graph convolutional
layer [Chen et al., He et al.], use
graph attention networks to
recognize meaningful user-item
interactions at higher-grained level
[Wang et al. (2020), Tao et al.]
2021-2022
Self-supervised and
contrastive learning
[Wu et al., Yu et al.]
2021-2022
Simplify the message-
passing formulation [Mao
et al., Peng et al., Shen et
al.]
2021-2022
Explore other latent
spaces [Shen et al., Sun
et al., Zhang et al.]
2022
Exploit hypergraphs [Wei
et al., Xia et al.]
2022
Use graph attention networks to
recognize meaningful user-item
interactions at finer-grained
[Zhang et al.] level
2017-2018
Pioneer approaches
proposing GCN-based
aggregation methods
[van den Berg et al.,
Ying et al.]
2019
Explore inter-dependencies between nodes
and their neighbors [Wang et al. (2019a)],
use graph attention networks to recognize
meaningful user-item interactions at
higher-grained level [Wang et al.
(2019b)]
5
Challenging the Myth of Graph Collaborative Filtering: a Reasoned and Reproducibility-driven Analysis
The 17th ACM Conference on Recommender Systems (Singapore, 18-22 September 2023)

6. Challenging the Myth of Graph Collaborative Filtering: a Reasoned and Reproducibility-driven Analysis
The 17th ACM Conference on Recommender Systems (Singapore, 18-22 September 2023)
● Reproducibility in machine learning research is the cutting-edge task involving the replication
of experimental results under the same share settings [Bellogín and Said, Anelli et al. (2021a-
2022), Ferrari Dacrema et al. (2019-2021), Sun et al.]
● In graph collaborative filtering, reproducibility is not always feasible since novel approaches
usually tend to
○ copy and paste previous results from the baselines
○ do not provide full details about the experimental settings
● What the research community should seek to
○ provide more detailed descriptions of the experimental settings
○ establish standard evaluation metrics and experimental protocols
Reproducibility and graph collaborative filtering
6

7. Challenging the Myth of Graph Collaborative Filtering: a Reasoned and Reproducibility-driven Analysis
The 17th ACM Conference on Recommender Systems (Singapore, 18-22 September 2023)
● RQ1. Is the state-of-the-art (i.e., the six most important papers) of graph collaborative filtering
(graph CF) replicable?
● RQ2. How does the state-of-the-art of graph CF position with respect to classic CF state-of-the-art?
● RQ3. How does the state-of-the-art of graph CF perform on datasets from different domains and
with different topological aspects, not commonly adopted for graph CF recommendation?
● RQ4. What information (or lack of it) impacts the performance of the graph CF methods across the
various datasets?
Research questions
7

8. Background and
reproducibility analysis
8

9. Challenging the Myth of Graph Collaborative Filtering: a Reasoned and Reproducibility-driven Analysis
The 17th ACM Conference on Recommender Systems (Singapore, 18-22 September 2023)
Background notions
i2
i3
u1
i1
u2
1 1 1
1 1 0
R
0 0 1 1 1
0 0 1 1 0
1 1 0 0 0
1 1 0 0 0
1 0 0 0 0
A
User-item interaction matrix
Adjacency matrix
Bipartite and undirected
user-item graph
U I
u1 u2
i3
i2
i1
9

10. Challenging the Myth of Graph Collaborative Filtering: a Reasoned and Reproducibility-driven Analysis
The 17th ACM Conference on Recommender Systems (Singapore, 18-22 September 2023)
Selected graph-based recommender systems
10
Model Venue Year Strategy
NGCF SIGIR 2019
• Pioneer approach in graph CF
• Inter-dependencies among ego and neighbor nodes
DGCF SIGIR 2020
• Disentangles users’ and items’ into intents and weights their importance
• Updates graph structure according to those learned intents
LightGCN SIGIR 2020
• Lightens the graph convolutional layer
• Removes feature transformation and non-linearities
SGL SIGIR 2021
• Brings self-supervised and contrastive learning to recommendation
• Learns multiple node views through node/edge dropout and random walk
UltraGCN CIKM 2021
• Approximates infinite propagation layers through a constraint loss and negative sampling
• Explores item-item connections
GFCF CIKM 2021
• Questions graph convolution in recommendation through graph signal processing
• Proposes a strong close-form algorithm

11. Challenging the Myth of Graph Collaborative Filtering: a Reasoned and Reproducibility-driven Analysis
The 17th ACM Conference on Recommender Systems (Singapore, 18-22 September 2023)
● Most of the approaches (apart
from UltraGCN) are compared
against a small subset of
classical CF solutions
● The recent literature has raised
concerns about usually-
untested strong CF baselines
[Anelli et al. (2021a-2022),
Ferrari Dacrema et al. (2019-
2021), Zhu et al.]
Analysis on reported baselines
11
Families Baselines
Models
NGCF [71] DGCF [73] LightGCN [28] SGL [78] UltraGCN [47] GFCF [59]
Used as graph CF baseline in (2021 — present)
[10, 13, 32, 62, 77, 84] [19, 39, 46, 74, 75, 92] [40, 54, 78, 82, 88, 89] [22, 46, 77, 82, 85, 93] [17, 24, 42, 48, 95, 96] [4, 5, 41, 50, 80, 96]
Classic CF
MF-BPR [55] 3 3 3
NeuMF [29] 3
CMN [18] 3
MacridVAE [44] 3
Mult-VAE [38] 3 3 3
DNN+SSL [86] 3
ENMF [11] 3
CML [30] 3
DeepWalk [52] 3
LINE [66] 3
Node2Vec [25] 3
NBPO [91] 3

12. Challenging the Myth of Graph Collaborative Filtering: a Reasoned and Reproducibility-driven Analysis
The 17th ACM Conference on Recommender Systems (Singapore, 18-22 September 2023)
● Conversely, most of the approaches
are compared against graph CF
solutions
● Orange ticks indicate that no
extensive comparison among the
selected baselines is performed (for
chronological reasons)
Analysis on reported baselines (cont.)
12
Families Baselines
Models
NGCF [71] DGCF [73] LightGCN [28] SGL [78] UltraGCN [47] GFCF [59]
Used as graph CF baseline in (2021 — present)
[10, 13, 32, 62, 77, 84] [19, 39, 46, 74, 75, 92] [40, 54, 78, 82, 88, 89] [22, 46, 77, 82, 85, 93] [17, 24, 42, 48, 95, 96] [4, 5, 41, 50, 80, 96]
Graph CF
HOP-Rec [83] 3
GC-MC [68] 3 3
PinSage [87] 3
NGCF [71] 3 3 3 3 3
DisenGCN [43] 3
GRMF [53] 3 3
GRMF-Norm [28] 3 3
NIA-GCN [64] 3
LightGCN [28] 3 3 3
DGCF [73] 3
LR-GCCF [14] 3
SCF [94] 3
BGCF [63] 3
LCFN [90] 3

13. Challenging the Myth of Graph Collaborative Filtering: a Reasoned and Reproducibility-driven Analysis
The 17th ACM Conference on Recommender Systems (Singapore, 18-22 September 2023)
Analysis on reported datasets
● Only a limited subset of shared
recommendation datasets
● We include novel, never-investigated
datasets
13
Models Gowalla Yelp 2018 Amazon Book Alibaba-iFashion Movielens 1M Amazon Electronics Amazon CDs
NGCF 3 3 3
DGCF 3 3 3
LightGCN 3 3 3
SGL 3 3 3
UltraGCN 3 3 3 3 3 3
GFCF 3 3 3

14. Challenging the Myth of Graph Collaborative Filtering: a Reasoned and Reproducibility-driven Analysis
The 17th ACM Conference on Recommender Systems (Singapore, 18-22 September 2023)
Analysis on experimental comparison
● NGCF train all baselines from scratch
● DGCF reports the results directly from the NGCF paper for the shared baselines
● LightGCN, SGL, and UltraGCN copy and paste from the original papers
● GFCF reproduce the results from LightGCN as the baselines are exactly the same
● Some authors are shared across such works
What we have done
● Re-implement from scratch all baselines by carefully following the original works
● Train/evaluate them within Elliot [Anelli et al. (2021b), Malitesta et al. (2023a)]
● Our goal is to provide a fair and repeatable experimental environment
● Use the same hyper-parameter settings as reported in the original papers and codes
14

15. Replication of prior results (RQ1)
15

16. Challenging the Myth of Graph Collaborative Filtering: a Reasoned and Reproducibility-driven Analysis
The 17th ACM Conference on Recommender Systems (Singapore, 18-22 September 2023)
● All approaches (except for SGL) use the same datasets filtering and splitting (80/20 hold-out splitting
user-wise)
● 10% of the training is left for validation for the tuning of hyper-parameters (no indication in the
papers and/or codes)
● All unrated items as evaluation protocol
● Evaluation through the Recall@20 and nDCG@20 (Recall@20 as validation metric)
● The best settings of hyper-parameters are usually shared in the paper and/or code
Settings
16

17. Challenging the Myth of Graph Collaborative Filtering: a Reasoned and Reproducibility-driven Analysis
The 17th ACM Conference on Recommender Systems (Singapore, 18-22 September 2023)
Results
● The most significant performance shift is in the order
of 10-3
● GFCF is the best replicated one (no random
initialization of model weights)
● NGCF and DGCF rarely achieve 10-4
because of the
random initializations and stochastic learning
processes involved
● Replicability is ensured and the copy-paste practise
did not hurt the results
17
Datasets Models Ours Original Performance Shift
Recall nDCG Recall nDCG Recall nDCG
Gowalla
NGCF 0.1556 0.1320 0.1569 0.1327 1.3 · 10 03 7 · 10 04
DGCF 0.1736 0.1477 0.1794 0.1521 5.8 · 10 03 4.4 · 10 03
LightGCN 0.1826 0.1545 0.1830 0.1554 4 · 10 04 9 · 10 04
SGL* — — — — — —
UltraGCN 0.1863 0.1580 0.1862 0.1580 +1 · 10 04 0
GFCF 0.1849 0.1518 0.1849 0.1518 0 0
Yelp 2018
NGCF 0.0556 0.0452 0.0579 0.0477 2.3 · 10 03 2.5 · 10 03
DGCF 0.0621 0.0505 0.0640 0.0522 1.9 · 10 03 1.7 · 10 03
LightGCN 0.0629 0.0516 0.0649 0.0530 2 · 10 03 1.4 · 10 03
SGL 0.0669 0.0552 0.0675 0.0555 6 · 10 04 3 · 10 04
UltraGCN 0.0672 0.0553 0.0683 0.0561 1.1 · 10 03 8 · 10 04
GFCF 0.0697 0.0571 0.0697 0.0571 0 0
Amazon Book
NGCF 0.0319 0.0246 0.0337 0.0261 1.8 · 10 03 1.5 · 10 03
DGCF 0.0384 0.0295 0.0399 0.0308 1.5 · 10 03 1.3 · 10 03
LightGCN 0.0419 0.0323 0.0411 0.0315 +8 · 10 04 +8 · 10 04
SGL 0.0474 0.0372 0.0478 0.0379 4 · 10 04 7 · 10 04
UltraGCN 0.0688 0.0561 0.0681 0.0556 +7 · 10 04 +5 · 10 04
GFCF 0.0710 0.0584 0.0710 0.0584 0 0
*Results are not provided since SGL was not originally trained and tested on Gowalla.

18. Benchmarking graph CF approaches using
alternative baselines (RQ2)
18

19. Challenging the Myth of Graph Collaborative Filtering: a Reasoned and Reproducibility-driven Analysis
The 17th ACM Conference on Recommender Systems (Singapore, 18-22 September 2023)
● Expand the investigation to four classic CF recommender systems: UserkNN, ItemkNN, RP3β, EASER
[Ferrari Dacrema et al. (2019), Anelli et al. (2022)]
● Consider two unpersonalized approaches (MostPop and Random)
● Follow the exact same 80/20 train/test splitting, and retain our version of the 10% of the training
as validation
● Use Tree-structured Parzen Estimator (with 20 exploration) [Bergstra et al.]
● Recall@20 is used as validation metric
Settings
19
Challenging the Myth of Graph Collaborative Filtering: a Reasoned and Reproducibility-driven Analysis
The 17th ACM Conference on Recommender Systems (Singapore, 18-22 September 2023)

20. Challenging the Myth of Graph Collaborative Filtering: a Reasoned and Reproducibility-driven Analysis
The 17th ACM Conference on Recommender Systems (Singapore, 18-22 September 2023)
Results
● Neither MostPop nor Random get acceptable results:
popularity bias is not present in the datasets or was
removed (see later)
● Some of the classic CF approaches reach better
performance than some graph CF baselines, and on Yelp
2018 and Amazon Book they reach best or second-to-best
performance
20
Families Models Gowalla Yelp 2018 Amazon Book
Recall nDCG Recall nDCG Recall nDCG
Reference
MostPop 0.0416 0.0316 0.0125 0.0101 0.0051 0.0044
Random 0.0005 0.0003 0.0005 0.0004 0.0002 0.0002
Classic CF
UserkNN 0.1685 0.1370 0.0630 0.0528 0.0582 0.0477
ItemkNN 0.1409 0.1165 0.0610 0.0507 0.0634 0.0524
RP3 0.1829 0.1520 0.0671 0.0559 0.0683 0.0565
EASER* 0.1661 0.1384 0.0655 0.0552 0.0710 0.0567
Graph CF
NGCF 0.1556 0.1320 0.0556 0.0452 0.0319 0.0246
DGCF 0.1736 0.1477 0.0621 0.0505 0.0384 0.0295
LightGCN 0.1826 0.1545 0.0629 0.0516 0.0419 0.0323
SGL — — 0.0669 0.0552 0.0474 0.0372
UltraGCN 0.1863 0.1580 0.0672 0.0553 0.0688 0.0561
GFCF 0.1849 0.1518 0.0697 0.0571 0.0710 0.0584
*Results for EASER on Amazon Book are taken from BARS Benchmark.

21. Extending the experimental comparison to new
datasets (RQ3 - RQ4)
21

22. Challenging the Myth of Graph Collaborative Filtering: a Reasoned and Reproducibility-driven Analysis
The 17th ACM Conference on Recommender Systems (Singapore, 18-22 September 2023)
● Two novel datasets: Allrecipes and BookCrossing with
discordant characteristics compared to the other datasets
● Allrecipes:
○ users are more numerous than items
○ much lower average user and item degrees
● BookCrossing:
○ lowest ratio between users and items
○ much higher density than the other datasets
● Useful to assess the performance in different (and never-
explored) topological settings
● Use the same experimental setting from RQ2 but with
validation set (10% of the training set)
Settings
22
Statistics Gowalla Yelp 2018 Amazon Book Allrecipes BookCrossing
Users 29,858 31,668 52,643 10,084 6,754
Items 40,981 38,048 91,599 8,407 13,670
Edges 810,128 1,237,259 2,380,730 80,540 234,762
Density 0.0007 0.0010 0.0005 0.0010 0.0025
Avg. Deg. (U ) 27.1327 39.0697 45.2241 7.9869 34.7590
Avg. Deg. (I ) 19.7684 32.5184 25.9908 9.5801 17.1735

23. Challenging the Myth of Graph Collaborative Filtering: a Reasoned and Reproducibility-driven Analysis
The 17th ACM Conference on Recommender Systems (Singapore, 18-22 September 2023)
Results
● Classic CF methods are very competitive
● Especially on BookCrossing, the classic CF baselines are the
top-performing approaches
● Only UltraGCN and LightGCN keep their performance as
observed in the previous datasets
● For the other graph-based ones, the performance
significantly drops
23
Families Models Allrecipes BookCrossing
Recall nDCG Recall nDCG
Reference
MostPop 0.0472 0.0242 0.0352 0.0319
Random 0.0024 0.0010 0.0013 0.0011
Classic CF
UserkNN 0.0339 0.0188 0.0871 0.0769
ItemkNN 0.0326 0.0180 0.0779 0.0739
RP3 0.0170 0.0089 0.0941 0.0821
EASER 0.0351 0.0192 0.0925 0.0847
Graph CF
NGCF 0.0291 0.0144 0.0670 0.0546
DGCF 0.0448 0.0234 0.0643 0.0543
LightGCN 0.0459 0.0236 0.0803 0.0660
SGL 0.0365 0.0192 0.0716 0.0600
UltraGCN 0.0475 0.0248 0.0800 0.0651
GFCF 0.0101 0.0051 0.0819 0.0712

24. Challenging the Myth of Graph Collaborative Filtering: a Reasoned and Reproducibility-driven Analysis
The 17th ACM Conference on Recommender Systems (Singapore, 18-22 September 2023)
Discussion (graph-based models’ ranking)
● UltraGCN and GFCF are the two best-performing
approaches
● All the other approaches rank according to the
chronological order
● On Allrecipes and BookCrossing
○ UltraGCN preserves its role of leading approach
○ GFCF and DGCF performance is very
fluctuating
○ LightGCN is in the top positions and surpasses
other models which should ideally outperform
it (e.g., SGL)
○ NGCF poor performance is confirmed
24
Metric Gowalla Yelp 2018 Amazon Book Allrecipes BookCrossing
Recall
1. UltraGCN (+19.73%) GFCF (+25.36%) GFCF (+122.57%) UltraGCN (+370.30%) GFCF (+27.37%)
2. GFCF (+18.83%) UltraGCN (+20.86%) UltraGCN (+115.67%) LightGCN (+354.46%) LightGCN (+24.88%)
3. LightGCN (+17.35%) SGL (+20.32%) SGL (+48.59%) DGCF (+343.56%) UltraGCN (+24.42%)
4. DGCF (+11.57%) LightGCN (+13.13%) LightGCN (+31.35%) SGL (+261.39%) SGL (+11.35%)
5. NGCF ( — ) DGCF (+11.69%) DGCF (+20.38%) NGCF (+188.12%) NGCF (+4.20%)
6. SGL* ( — ) NGCF ( — ) NGCF ( — ) GFCF ( — ) DGCF ( — )
nDCG
1. UltraGCN (+19.70%) GFCF (+26.33%) GFCF (+137.40%) UltraGCN (+386.27%) GFCF (+31.12%)
2. LightGCN (+17.05%) UltraGCN (+22.35%) UltraGCN (+128.05%) LightGCN (+362.75%) LightGCN (+21.55%)
3. GFCF (+15.00%) SGL (+22.12%) SGL (+51.22%) DGCF (+358.82%) UltraGCN (+19.89%)
4. DGCF (+11.89%) LightGCN (+14.16%) LightGCN (+31.30%) SGL (+276.47%) SGL (+10.50%)
5. NGCF ( — ) DGCF (+11.73%) DGCF (+19.92%) NGCF (+182.35%) NGCF (+0.55%)
6. SGL* ( — ) NGCF ( — ) NGCF ( — ) GFCF ( — ) DGCF ( — )
*SGL is not classifiable on the Gowalla dataset as results were not calculated in the original paper.

25. Challenging the Myth of Graph Collaborative Filtering: a Reasoned and Reproducibility-driven Analysis
The 17th ACM Conference on Recommender Systems (Singapore, 18-22 September 2023)
Discussion (analysis on the node degree)
● We reinterpret node degree as information flow from
neighbor nodes to the ego nodes after multiple hops
● Only users as ending nodes because accuracy metrics are
calculated user-wise
● Information flow at 1, 2, and 3 hops:
information after 1-hop
column vector
25

26. Challenging the Myth of Graph Collaborative Filtering: a Reasoned and Reproducibility-driven Analysis
The 17th ACM Conference on Recommender Systems (Singapore, 18-22 September 2023)
Analysis on the node degree (1-hop)
Indication of the activeness of users on the platform
average performance
(nDCG@20)
user quartiles over
information values
performance
improvement
● The 4th quartile is favoured with respect to the other ones
● The trend is even more evident on GFCF
26

27. Challenging the Myth of Graph Collaborative Filtering: a Reasoned and Reproducibility-driven Analysis
The 17th ACM Conference on Recommender Systems (Singapore, 18-22 September 2023)
Analysis on the node degree (2-hop)
Indication of the influence of items’ popularity on users
● Models favour the warm users who enjoyed popular items over the cold users who enjoyed niche
items
● On Allrecipes, UltraGCN, DGCF, and LightGCN show less discriminatory behavior across
quartiles; SGL and NGCF show a higher slope that is comparable to classic CF methods; GFCF
behavior is even more accentuated than the 1-hop setting
● On BookCrossing, the trend is almost aligned across all models
27

28. Challenging the Myth of Graph Collaborative Filtering: a Reasoned and Reproducibility-driven Analysis
The 17th ACM Conference on Recommender Systems (Singapore, 18-22 September 2023)
Analysis on the node degree (3-hop)
Indication of the influence of co-interacting users’ activeness on users
● On Allrecipes, UltraGCN, DGCF, and LightGCN exhibit more
consistency across quartiles, while NGCF, SGL, and GFCF have a
more disparate range of results
● On BookCrossing, the information at the 3-hop is not providing
more insights than the 2-hop
28

29. Conclusion and future work
29

30. Challenging the Myth of Graph Collaborative Filtering: a Reasoned and Reproducibility-driven Analysis
The 17th ACM Conference on Recommender Systems (Singapore, 18-22 September 2023)
Conclusion
● Replicate the results of six state-of-the-art graph CF methods
● We include other state-of-the-art approaches and other (unexplored) datasets
● The topological graph characteristics (i.e., node degree) may impact the performance
● This happens especially for the information flow at 2-hop (i.e., user activeness + item popularity)
Future work
● Further investigation into diversity and fairness of graph CF approaches
● Analyze the impact of other topological graph characteristics on the performance (currently on arXiv
[Malitesta et al. (2023b)])
30

31. Challenging the Myth of Graph Collaborative Filtering: a Reasoned and Reproducibility-driven Analysis
The 17th ACM Conference on Recommender Systems (Singapore, 18-22 September 2023)
Useful resources
31

32. Challenging the Myth of Graph Collaborative Filtering: a Reasoned and Reproducibility-driven Analysis
The 17th ACM Conference on Recommender Systems (Singapore, 18-22 September 2023)
A Topology-aware Analysis of Graph Collaborative Filtering
32
[Malitesta et al. (2023b)]

33. Challenging the Myth of Graph Collaborative Filtering: a Reasoned and Reproducibility-driven Analysis
The 17th ACM Conference on Recommender Systems (Singapore, 18-22 September 2023)
[van den Berg et al.] 2017. Graph convolutional matrix completion. CoRR abs/1706.02263.
[Ying et al.] 2018. Graph Convolutional Neural Networks for Web-Scale Recommender Systems. In KDD. ACM, 974–983.
[Wang et al. (2019a)] 2019. Neural Graph Collaborative Filtering. In SIGIR. ACM, 165–174.
[Wang et al. (2019b)] 2019. KGAT: Knowledge Graph Attention Network for Recommendation. In KDD. ACM, 950–958.
[Chen et al.] 2020. Revisiting Graph Based Collaborative Filtering: A Linear Residual Graph Convolutional Network Approach. In AAAI. AAAI Press, 27–34.
[He et al.] 2020. LightGCN: Simplifying and Powering Graph Convolution Network for Recommendation. In SIGIR. ACM, 639–648.
[Wang et al. (2020)] 2020. Disentangled Graph Collaborative Filtering. In SIGIR. ACM, 1001–1010.
[Tao et al.] 2020. MGAT: Multimodal Graph Attention Network for Recommendation. Inf. Process. Manag. 57, 5 (2020), 102277.
[Wu et al.] 2021. Self-supervised Graph Learning for Recommendation. In SIGIR. ACM, 726–735.
[Yu et al.] 2022. Are Graph Augmentations Necessary?: Simple Graph Contrastive Learning for Recommendation. In SIGIR. ACM, 1294–1303.
[Mao et al.] 2021. UltraGCN: Ultra Simplification of Graph Convolutional Networks for Recommendation. In CIKM. ACM, 1253–1262.
[Peng et al.] 2022. SVD-GCN: A Simplified Graph Convolution Paradigm for Recommendation. In CIKM. ACM, 1625–1634.
[Shen et al.] 2021. How Powerful is Graph Convolution for Recommendation?. In CIKM. ACM, 1619–1629.
[Sun et al.] 2021. HGCF: Hyperbolic Graph Convolution Networks for Collaborative Filtering. In WWW. ACM / IW3C2, 593–601.
[Zhang et al.] 2022. Geometric Disentangled Collaborative Filtering. In SIGIR. ACM, 80–90.
[Wei et al.] 2022. Dynamic Hypergraph Learning for Collaborative Filtering. In CIKM. ACM, 2108–2117.
[Xia et al.] 2022. Hypergraph Contrastive Collaborative Filtering. In SIGIR. ACM, 70–79.
[Bellogín and Said] 2021. Improving accountability in recommender systems research through reproducibility. User Model. User Adapt. Interact. 31, 5 (2021), 941–977.
[Anelli et al. (2021a)] 2021. Reenvisioning the comparison between Neural Collaborative Filtering and Matrix Factorization. In RecSys. ACM, 521–529.
[Anelli et al. (2022)] 2022. Top-N Recommendation Algorithms: A Quest for the State-of-the-Art. In UMAP. ACM, 121–131.
References 1/2
33

34. Challenging the Myth of Graph Collaborative Filtering: a Reasoned and Reproducibility-driven Analysis
The 17th ACM Conference on Recommender Systems (Singapore, 18-22 September 2023)
[Ferrari Dacrema et al. (2019)] 2019. Are we really making much progress? A worrying analysis of recent neural recommendation approaches. In RecSys. ACM, 101–109.
[Ferrari Dacrema et al. (2021)] 2021. A Troubling Analysis of Reproducibility and Progress in Recommender Systems Research. ACM Trans. Inf. Syst. 39, 2 (2021), 20:1–
20:49.
[Sun et al.] 2020. Are We Evaluating Rigorously? Benchmarking Recommendation for Reproducible Evaluation and Fair Comparison. In RecSys. ACM, 23–32.
[Zhu et al.] 2022. BARS: Towards Open Benchmarking for Recommender Systems. In SIGIR. ACM, 2912–2923.
[Anelli et al. (2021b)] 2021. Elliot: A Comprehensive and Rigorous Framework for Reproducible Recommender Systems Evaluation. In SIGIR. ACM, 2405–2414.
[Malitesta et al. (2023a)] 2023. An Out-of-the-Box Application for Reproducible Graph Collaborative Filtering extending the Elliot Framework. In UMAP (Adjunct
Publication). ACM, 12–15.
[Bergstra et al.] 2011. Algorithms for Hyper-Parameter Optimization. In NIPS. 2546–2554.
[Malitesta et al. (2023b)] 2023. A Topology-aware Analysis of Graph Collaborative Filtering. CoRR abs/2308.10778.
References 2/2
34